Application of a Machine Learning Method for Prediction of Urban Neighborhood-Scale Air Pollution

Urban air pollution has aroused growing attention due to its associated adverse health effects. A model which could promptly predict urban air quality with considerable accuracy is, therefore, important and will benefit the development of smart cities. However, only a computational fluid dynamics (CFD) model could better resolve the dispersion behavior within an urban canyon layer. A machine learning (ML) model using the Artificial Neural Network (ANN) approach was formulated in the current study to investigate vehicle-derived airborne particulate (PM₁₀) dispersion within a compact high-rise-built environment. Various measured meteorological parameters and PM₁₀ concentrations were adopted as the model inputs to train the ANN model. A building-resolved CFD model under the same environmental settings was also set up to compare its model performance with the ANN model. Our results showed that the ANN model exhibited promising performance (r = 0.82, fractional bias = 0.002) when comparing the > 1000 h PM₁₀ measurements. When comparing the diurnal hourly measured PM₁₀ variations in a clear-sky day, both the ANN and CFD models performed well (r > 0.8). The good performance of the CFD model relied on the knowledge of the in situ diurnal traffic profile, the adoption of suitable mobile source emission factor(s) (e.g., from MOBILE 6 and COPERT4), and the use of urban thermal and dynamical variables to capture PM₁₀ variations in both neutral and unstable atmospheric conditions. These requirements/constraints make it impractical for daily operation. On the contrary, the ML (ANN) model adopted here is free from these constraints and is fast (less than 0.1% computational time relative to the CFD model). These results demonstrate that the ANN model is a superior option for a smart city application.